Weakly supervised segmentation models as explainable radiological classifiers for lung tumour detection on CT images.

Area of Science:

• Radiological imaging
• Medical image analysis
• Deep learning for healthcare

Background:

• Interpretability is crucial for reliable convolutional neural network (CNN) image classifiers in radiology.
• Existing methods often require detailed annotations, increasing workload.

Purpose of the Study:

• To develop a weakly supervised segmentation model for explainable object detection in radiological images.
• To train a model using only image-level labels, reducing annotation burden.

Main Methods:

• A weakly supervised U-Net architecture (WSUnet) was trained for lung tumor segmentation using image-level labels.
• WSUnet generates voxel probability maps and uses global max-pooling for image-level prediction.
• WSUnet's performance was compared against traditional interpretation techniques using CT data from multiple institutions.

Main Results:

• WSUnet accurately localized tumors with high precision (0.77-0.78) and competitive dice scores (0.33-0.43) without voxel-level training labels.
• WSUnet's voxel-level discrimination outperformed comparator methods, evidenced by higher area under the precision-recall curve (AUPR).
• Clinicians showed a strong preference for WSUnet's predictions (72% preference rate).

Conclusions:

• Weakly supervised segmentation is a viable method for creating explainable object detection models in medical imaging.
• WSUnet provides inherently explainable predictions at the voxel level, enhancing model interpretability.
• This approach reduces the need for extensive manual annotation, streamlining the development of medical image classifiers.